Reasoning Errors Have a Region and a Direction in the Residual-Stream Trajectory of LLMs
Hamed Damirchi, Ignacio Meza De la Jara, Damith Ranasinghe, Yuhang Liu, Javen Shi
Why It Matters
What makes this one worth your time
Understanding and improving reasoning in language models is crucial for tasks requiring verifiable reasoning, impacting AI reliability and trustworthiness.
A novel three-stream detector enhances reasoning error detection in language models by integrating motion and location views.
Summary
The paper proposes a three-stream detector for identifying reasoning errors in language models by combining motion with two restricted views of location, improving reasoning accuracy on benchmarks and outperforming existing methods.
Key contributions
- Introduction of a three-stream detector combining motion and location views.
- Improved reasoning accuracy on unseen benchmarks by up to 12% over state-of-the-art methods.
- Demonstrated applicability to factual completion and verification tasks.
Notable insights
- Combining motion with restricted location views can enhance reasoning error detection.
- State-conditioned motion provides a more accurate signal for reasoning validity than static states or decontextualized trajectories.
Possible limitations
- Not stated in the abstract
Abstract
arXiv:2608.05660v1 Announce Type: cross Abstract: As language models are increasingly used for tasks that require verifiable reasoning, reliably distinguishing sound reasoning from flawed reasoning has become an important practical problem. Recent trajectory-based methods seek this signal in layerwise residual-stream displacements, which capture how representations change while attenuating some stable, token-specific information. However, displacement omits the state from which an update originates, whereas restoring the full state risks reintroducing shortcut-prone information. We identify this trade-off and propose a three-stream detector that combines motion with two restricted views of location. A coarse region reader based on vector quantization and a fine direction reader over normalized multi-layer states. This design restores enough state context to interpret the motion without returning to full-state probing. On reasoning benchmarks unseen during training, our method improves selection accuracy by up to 12% over the displacement-only state of the art and 21% over single-layer probing baselines. Although trained only on reasoning benchmarks, it also reads factual completion and fact verification, ahead of every detector we compare against, which places the signal on correctness rather than on a kind of reasoning. Ablations further show that motion, region, and direction provide complementary signals. These results suggest that reasoning validity is better read from state-conditioned motion than from either static states or decontextualized trajectories alone.